Devin Bot built an Android app on its own in 44 days: 1 134 PRs from 183 messages
Cognition has launched Devin Bots, persistent AI agents that receive a goal rather than a task. Instead of "port this feature to Android," the instruction is "keep Android at parity with iOS." In 44 days, the first bot took the Devin Android app to open testing on its own. It merged 1 134 PRs (code changes that were accepted into the project), and a human sent it only 183 messages. In practice, instead of managing tickets, the engineer is left to set the goal and the success criterion.
AI-processed from Windsurf Blog; edited by Hamidun News
Cognition has launched Devin Bots, persistent AI agents that are assigned an entire goal rather than a single task. The difference is simple. You tell an ordinary coding agent, "port this feature to Android," and it does exactly that.
A human still has to find the next problem, write a new ticket and check the result. A Devin Bot gets the instruction "keep Android at parity with iOS" and then decides for itself what to build. The most striking example is the Devin Android app: the bot took it to open testing on Google Play in 44 days.
In that time it merged 1 134 pull requests (a PR is a proposed code change that passed checks and landed in the project), while a human sent it just 183 messages. If you're tired of being a "manager" for your own AI agent, this is for you.
What is a Devin Bot
A Bot is a persistent agent with its own memory and tools. It can run several Devin sessions at once: it investigates a problem, hands out work, evaluates the results and picks what to take on next. Context carries over between sessions. The bot comes to a human only with decisions that require human judgment, and these pile up in a Needs You inbox. You can talk to bots in Slack, iMessage or directly in the Devin app.
Cognition gives several example pairs: instead of "fix the layout shift," "make sure the web app doesn't jitter when things change"; instead of "fix the onboarding bug," "make sure new users get through setup."
Three case studies from the company itself
The first is Android. Cognition had a team of iOS engineers and not a single Android developer. On September 28, 2026, engineer Samyok created an "Android Bot" and told it to run the product and build the app autonomously, starting by identifying blockers.
The bot filed tickets, launched a Devin session for each one and merged PRs that passed CI and Devin Review. The result: 1 134 merged PRs, 344 651 lines of Kotlin, 2 841 screenshot tests and 1 217 closed tickets. Samyok wrote roughly one message per six PRs, and on 22 of the 44 days he wrote nothing at all.
Most messages were answers to questions like "does this PR need a security review," and occasionally encouragement like "be more ambitious."
The second is Layout Bot. Page elements "jump" when content arrives late, especially while an agent's response is streaming. This is hard to catch in code review: it depends on request timing and on whether the user has visited before.
Every three hours, the bot builds the production version of main, brings up the entire stack on its own machine and records a first visit, a repeat visit and loads with deliberately delayed requests. In one case, a row in the sidebar appeared 0.9 seconds later than the others and shifted the list.
The bot found the PR that introduced the bug and opened a fix with a before-and-after video. In three weeks it merged 159 PRs and cut lag by 70%. Overall, Cognition merges more than 2 000 PRs per week.
The third is a game studio. Jenny Xu is the only human left in a mobile studio where about 20 bots each own one area: development, Unity, releases, UI, iOS, Android and testing on real devices. Above them sits Super Jenny Bot, the only one she talks to. Four times a day it collects questions into a single Slack thread, and she answers with emoji. The studio grew from 50 to 1 900 PRs per week and ships a release every week.
What is left for humans, and the caveats
Cognition says outright that the work doesn't disappear but shifts. The engineer defines what counts as success. A bot that speeds up CI has to know which checks cannot be removed and whether it can spend more on compute. Feedback is also needed: tests, benchmarks and user scenarios, and in production, usage metrics and customer reviews. Humans also set the limits of autonomy: which systems the bot can access, what it can change, and when releases and spending require approval.
It is important to remember that the figures come from the company itself, and there is no independent verification. The quality of the 1 134 PRs and 344 thousand lines of Kotlin cannot be judged from the post. All we know is that they passed CI and Devin Review. The app is still in open beta. The announcement does not disclose the cost of tokens and compute, though a bot that builds and runs the entire stack every three hours is clearly not free. One of the customer testimonials says their bots merged about 1 600 PRs last week.
The practical takeaway: if you have useful work that perpetually loses out to urgent tasks, such as platform parity, regression monitoring or release support, try describing it as an area of responsibility with a measurable criterion. Without automated verification of the result, such a bot will simply churn out a lot of PRs.
- Nov 18, 2026Google finds AI agents that discover vulnerabilities and hack websites on their own
- Nov 18, 2026AI agent ShakeFix fixes a pharmacy app bug in 18 minutes at GitLab hackathon
- Nov 18, 2026Nvidia's data-center revenue jumps 110% to $107.6 billion: the AI pause didn't hurt
- Nov 18, 2026AI Finds Unknown V8 Vulnerability for $4.24 in 13 Minutes, After Failing for $350 in February
Key Questions Answered
What is a Devin Bot?
A Bot is a persistent agent with its own memory and tools. It can run several Devin sessions at once: it investigates a problem, hands out work, evaluates the results and picks what to take on next. Context carries over between sessions. The bot comes to a human only with decisions that require human judgment, and these pile up in a Needs You inbox. You can talk to bots in Slack, iMessage or directly in the Devin app.
What are the key details on Three case studies from the company itself?
The first is Android. Cognition had a team of iOS engineers and not a single Android developer. On September 28, 2026, engineer Samyok created an "Android Bot" and told it to run the product and build the app autonomously, starting by identifying blockers. The bot filed tickets, launched a Devin session for each one and merged PRs that passed CI and Devin Review. The result: 1 134 merged PRs, 344 651 lines of Kotlin, 2 841 screenshot tests and 1 217 closed tickets.
What is left for humans, and the caveats?
Cognition says outright that the work doesn't disappear but shifts. The engineer defines what counts as success. A bot that speeds up CI has to know which checks cannot be removed and whether it can spend more on compute. Feedback is also needed: tests, benchmarks and user scenarios, and in production, usage metrics and customer reviews. Humans also set the limits of autonomy: which systems the bot can access, what it can change, and when releases and spending require approval.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.